Your team, shipping real AI agents — by Friday.
A hands-on workshop run by Oleg Podsechin — the engineer behind Ghostlyn and ten years of Toughbyte. Cuts through the agent hype, maps it onto your workflows, and leaves your team with a working prototype they actually use on Monday.
Built by people who ship agents — not people who tweet about them.
Ghostlyn is model-agnostic — DeepSeek, Claude, OpenAI, open-weight, switch mid-run. Built for teams: shared skills, shared tools, shared memory.
Vercel AI under the hood. Progressive disclosure for skills. Sub-agent fan-out for parallel work. Structured memory tables, not just markdown.
Most workshops are generic. Ours starts with a discovery call. We build the workshop around the workflows you actually want to automate.
Six concrete things, not vibes.
How to define the name, description and parameters so the model actually picks the right tool — and how to spot when it won't.
Million-token windows fill fast. Compaction discipline, sub-agents, fresh chats per task, offloading state to files.
Bundle expertise into SKILL.md folders with progressive disclosure. When to skill, when to MCP, when to just prompt.
Grep, full-text, embeddings. Structured local tables vs unstructured files. Stopping hallucinations with grounded retrieval.
Private data + untrusted content + external comms = RCE. Permission modes, auto-mode monitors, blast-radius design.
When code-driving-LLM beats LLM-driving-code. Answer agents vs act agents. Human-in-the-loop drafting patterns.
Five modules. One working prototype.
The exact agenda is shaped by your discovery call. This is the spine we work from.
- M01Foundations — what an agent actually is
Model, harness, tools, context. The agentic loop. Why the model isn't the bottleneck anymore.
- M02Your LLM mental model
Next-token prediction, hallucination, memory-as-illusion. Sizing the right model for the job — proprietary vs open-weight cost math.
- M03Building blocks & tools
Function calling, web fetch, code execution, MCP. Designing tools the model can use without supervision.
- M04Architectures & safety
Skills with progressive disclosure, sub-agent fan-out, the lethal trifecta, permission modes, the auto-mode monitor pattern.
- M05Hands-on with Ghostlyn
Pick a real workflow from your team. Build it end-to-end on Ghostlyn against your data. Ship it.
By the end of the day, your team can…
- Run a sane discovery on which workflows are actually worth automating (and which are vanity demos).
- Build, deploy and iterate on a Ghostlyn agent against their own data.
- Tell when a problem wants a workflow vs an agent — and design accordingly.
- Spot the lethal trifecta in a proposed design before it ships.
- Manage context windows, skills and sub-agents like senior practitioners.
Teams who are past “ChatGPT for everyone”.
rolling out agents internally and tired of unmaintained demos.
rethinking the product workflow now that LLMs are in the loop.
automating repetitive, judgment-light parts of the funnel.
validating where AI actually changes the unit economics.
Oleg Podsechin
Founder of Toughbyte (10 years). Builder of Ghostlyn, a team-first AI work assistant. Built one of the earliest mobile multiplayer games — pre-home-computer. Has been writing software since before stack overflow had a stack to overflow. Spends his days shipping agents that companies actually use.
A working agent, in your team's hands, by the end of the day.
Tailored Ghostlyn Workshops from €1,950. One day, six to twelve people, your stack, your data, your workflow. Reply to a short discovery email — we take it from there.